Engineering leaders, security owners, and research peers, in their own words. Anyone here is happy to take a reference call.
He found the failure we had convinced ourselves was theoretical, then wrote the test that keeps it from coming back.
We hired Blake for two weeks of prompt injection testing and kept him for the quarter. On day three he refunded a fictional customer's invoice using nothing but text hidden in a help-center article one of our partners could edit. That demo ended an argument our team had been having for months, and the confirmation step shipped the following sprint.
Blake joined a platform team that was skeptical of outside security help and left with three of our engineers using his tooling by choice. His detections are still running a year later, untouched, which is the only measure of this work I trust.
We came to him with an architecture we thought was nearly done. He asked four questions, drew the trust boundaries we had skipped, and we redesigned the session model before a line of it reached production. Every finding arrived with a fix we could schedule against real sprints.
Two weeks after our policy gate went live it blocked an unsigned experimental checkpoint headed for production traffic. That is an incident that never happened, and it is the clearest return I can point to from any security engagement we have run.
I cite Blake's poisoning numbers in nearly every design review I run. He took a risk teams were writing off as academic and made it concrete enough to budget against, then published the dataset so the rest of us could check him.
Blake submitted the best-argued talk in the track and then spent the conference in the hallway helping other people's teams. He is the reviewer I send borderline agent-security papers to, because he will tell an author the uncomfortable thing kindly.
Our product engineers walked in expecting a lecture about jailbreaks and walked out with a logging checklist they had written themselves. Six months on, the phrase 'treat the tool result as untrusted' shows up unprompted in our pull request reviews.
He set up our exercise program and then deliberately made himself unnecessary — templates, a rotation schedule, and a debrief format the team still follows. Most consultants build dependence. Blake built capability and handed us the keys.